Small QOI fixes for transaction_kthread
cover Small QOI fixes for transaction_kthread
Following 4 patches make transaction_kthread code slitghly more user friendly. Namely, patch 1 convert from open-coded multiplicaiton by HZ to using msecs_to_jiffies helper. Patch 2 relies on the observation that the running transaction is obtained under trans_lock so an extra check can be removed. Patch 3/4 could possibly be squashed into 1 but the net effect is that the code is more intuitive when sleeping in case a lower interval than commit_trans has elapsed.
This has survived full xfstest run.
Nikolay Borisov (4):
btrfs: Use helpers to convert from seconds to jiffies in
transaction_kthread
btrfs: Remove redundant check
btrfs: Record delta directly in transaction_kthread
btrfs: Be smarter when sleeping in transaction_kthread
fs/btrfs/disk-io.c | 11 +++++------ 1 file changed, 5 insertions(+), 6 deletions(-)
--
2.17.1
PATCH 1/4 btrfs: Use helpers to convert from seconds to jiffies in transaction_kthread
The kernel provides easy to understand helpers to convert from human understandable units to the kernel-friendly 'jiffies'. So let's use those to make the code easier to understand. No functional changes.
--- fs/btrfs/disk-io.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
fs/btrfs/disk-io.c +2 -2
@@ -1721,7 +1721,7 @@ static int transaction_kthread(void *arg) do { cannot_commit = false; - delay = HZ * fs_info->commit_interval; + delay = msecs_to_jiffies(fs_info->commit_interval * 1000); mutex_lock(&fs_info->transaction_kthread_mutex); spin_lock(&fs_info->trans_lock);
@@ -1736,7 +1736,7 @@ static int transaction_kthread(void *arg) (now < cur->start_time || now - cur->start_time < fs_info->commit_interval)) {spin_unlock (&fs_info->trans_lock); - delay =HZ *5 ; + delay =msecs_to_jiffies (5000 ); goto sleep; } transid = cur->transid;
Signature · 1 line
2.17.1
Re: [PATCH 1/4] btrfs: Use helpers to convert from seconds to jiffies in transaction_kthread
Quoted 14 lines · Nikolay Borisov
The kernel provides easy to understand helpers to convert from human understandable units to the kernel-friendly 'jiffies'. So let's use those to make the code easier to understand. No functional changes.
Signed-off-by Nikolay Borisov <[email protected]>--- fs/btrfs/disk-io.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)fs/btrfs/disk-io.c +0
-0
fs/btrfs/disk-io.c static int transaction_kthread(void *arg) · line 1724
Quoted 5 lines
fs/btrfs/disk-io.c +1
-1@@ -1721,7 +1721,7 @@ static int transaction_kthread(void *arg) do { cannot_commit = false; - delay = HZ * fs_info->commit_interval; + delay = msecs_to_jiffies(fs_info->commit_interval * 1000);
Since we're now doing everything in msecs, why don't we just make sure ->commit_interval is set to msecs, that way we don't have to carry around the * 1000? If we still need the multiplication we should be using MSEC_PER_SEC. Thanks,
Josef
Re: [PATCH 1/4] btrfs: Use helpers to convert from seconds to jiffies in transaction_kthread
Quoted 10 lines · Josef Bacik
Quoted 9 lines · Nikolay Borisov
The kernel provides easy to understand helpers to convert from human understandable units to the kernel-friendly 'jiffies'. So let's use those to make the code easier to understand. No functional changes.
Signed-off-by Nikolay Borisov <[email protected]>---
fs/btrfs/disk-io.c| 4 ++--1 file changed, 2 insertions(+), 2 deletions(-)
Quoted 14 lines
fs/btrfs/disk-io.c static int transaction_kthread(void *arg) · line 1723
Quoted 9 lines
fs/btrfs/disk-io.c +1
-1@@ -1721,7 +1721,7 @@ static int transaction_kthread(void *arg) do { cannot_commit = false; - delay = HZ * fs_info->commit_interval; + delay = msecs_to_jiffies(fs_info->commit_interval * 1000);Since we're now doing everything in msecs, why don't we just make sure ->commit_interval is set to msecs, that way we don't have to carry around the * 1000? If we still need the multiplication we should be using MSEC_PER_SEC. Thanks,
Yes, as a matter of fact I intend on making commit_interval into jiffies, to completely eliminate the msecs_to_jiffies helpers here.
But that will be a follow on work once David merges the v2 of "Be smarter" patch.
Quoted 3 lines
Josef
Re: [PATCH 1/4] btrfs: Use helpers to convert from seconds to jiffies in transaction_kthread
Quoted 28 lines · Nikolay Borisov
Quoted 10 lines · Josef Bacik
Quoted 9 lines · Nikolay Borisov
The kernel provides easy to understand helpers to convert from human understandable units to the kernel-friendly 'jiffies'. So let's use those to make the code easier to understand. No functional changes.
Signed-off-by Nikolay Borisov <[email protected]>---
fs/btrfs/disk-io.c| 4 ++--1 file changed, 2 insertions(+), 2 deletions(-)Quoted 14 lines
fs/btrfs/disk-io.c static int transaction_kthread(void *arg) · line 1723
Quoted 9 lines
fs/btrfs/disk-io.c +1
-1@@ -1721,7 +1721,7 @@ static int transaction_kthread(void *arg) do { cannot_commit = false; - delay = HZ * fs_info->commit_interval; + delay = msecs_to_jiffies(fs_info->commit_interval * 1000);Since we're now doing everything in msecs, why don't we just make sure ->commit_interval is set to msecs, that way we don't have to carry around the * 1000? If we still need the multiplication we should be using MSEC_PER_SEC. Thanks,
Quoted 5 lines
Yes, as a matter of fact I intend on making
commit_intervalinto jiffies, to completely eliminate themsecs_to_jiffieshelpers here.But that will be a follow on work once David merges the v2 of "Be smarter" patch.
As the commit_interval is an internal value, jiffies is ok, the conversions happen happen for mount option set and print only. So ok from me.
PATCH 2/4 btrfs: Remove redundant check
In reply to Nikolay BorisovThe value obtained from ktime_get_seconds() is guaranteed to be monotonically increasing since it's taken from CLOCK_MONOTONIC. As transaction_kthread obtains a reference to the currently running transaction under holding btrfs_fs_info::trans_lock it's guaranteed to:
a) See an initialized 'cur', whose start_tim is guaranteed to be smaller than 'now'
or
b) Not obtain a 'cur' and simply go to sleep.
Given this remove the unnecessary check, if it sees now < cur->start_time this would imply there are far greated problems on the machine.
--- fs/btrfs/disk-io.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
fs/btrfs/disk-io.c +1 -2
@@ -1733,8 +1733,7 @@ static int transaction_kthread(void *arg) now =ktime_get_seconds ();if (cur->state <TRANS_STATE_COMMIT_START && - (now < cur->start_time || - now - cur->start_time < fs_info->commit_interval)) { + (now - cur->start_time < fs_info->commit_interval)) {spin_unlock (&fs_info->trans_lock); delay =msecs_to_jiffies (5000 ); goto sleep;
Signature · 1 line
2.17.1
PATCH 3/4 btrfs: Record delta directly in transaction_kthread
In reply to Nikolay BorisovRename 'now' to 'delta' and store there the delta between transaction start time and current time. This is in preparation for optimising the sleep logic in the next patch. No functional changes.
--- fs/btrfs/disk-io.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
fs/btrfs/disk-io.c +3 -3
@@ -1715,7 +1715,7 @@ static int transaction_kthread(void *arg)struct btrfs_trans_handle *trans;struct btrfs_transaction *cur;u64 transid; -time64_t now; +time64_t delta;unsigned long delay;bool cannot_commit;
@@ -1731,9 +1731,9 @@ static int transaction_kthread(void *arg) goto sleep; } - now =ktime_get_seconds (); + delta =ktime_get_seconds () - cur->start_time;if (cur->state <TRANS_STATE_COMMIT_START && - (now - cur->start_time < fs_info->commit_interval)) { + delta < fs_info->commit_interval) {spin_unlock (&fs_info->trans_lock); delay =msecs_to_jiffies (5000 ); goto sleep;
Signature · 1 line
2.17.1
PATCH 4/4 btrfs: Be smarter when sleeping in transaction_kthread
In reply to Nikolay BorisovIf transaction_kthread is woken up before btrfs_fs_info::commit_interval seconds have elapsed it will sleep for a fixed period of 5 seconds. This is not a problem per-se but is not accuaret, instead the code should sleep for an interval which guarantees on next wakeup commit_interval would have passed. Since time tracking is not accurate add 1 second to ensure next wake up would be after commit_interval.
--- fs/btrfs/disk-io.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
fs/btrfs/disk-io.c +1 -1
@@ -1735,7 +1735,7 @@ static int transaction_kthread(void *arg)if (cur->state <TRANS_STATE_COMMIT_START && delta < fs_info->commit_interval) {spin_unlock (&fs_info->trans_lock); - delay =msecs_to_jiffies (5000 ); + delay =msecs_to_jiffies ((1 +delta) *1000 ); goto sleep; } transid = cur->transid;
Signature · 1 line
2.17.1
Re: [PATCH 4/4] btrfs: Be smarter when sleeping in transaction_kthread
Quoted 18 lines · Nikolay Borisov
If
transaction_kthreadis woken up beforebtrfs_fs_info::commit_intervalseconds have elapsed it will sleep for a fixed period of 5 seconds. This is not a problem per-se but is not accuaret, instead the code should sleep for an interval which guarantees on next wakeupcommit_intervalwould have passed. Since time tracking is not accurate add 1 second to ensure next wake up would be aftercommit_interval.Signed-off-by Nikolay Borisov <[email protected]>--- fs/btrfs/disk-io.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)fs/btrfs/disk-io.c +0
-0
fs/btrfs/disk-io.c static int transaction_kthread(void *arg) · line 1738
Quoted 5 lines
fs/btrfs/disk-io.c +1
-1@@ -1735,7 +1735,7 @@ static int transaction_kthread(void *arg)if (cur->state <TRANS_STATE_COMMIT_START && delta < fs_info->commit_interval) {spin_unlock (&fs_info->trans_lock); - delay =msecs_to_jiffies (5000 ); + delay =msecs_to_jiffies ((1 +delta) *1000 );
This does not seem right. Delta is number of seconds since the transaction started, so we need to sleep for the remaining time. Which is commit_interval - delta.
Re: [PATCH 4/4] btrfs: Be smarter when sleeping in transaction_kthread
Quoted 14 lines · David Sterba
Quoted 13 lines · Nikolay Borisov
If
transaction_kthreadis woken up beforebtrfs_fs_info::commit_intervalseconds have elapsed it will sleep for a fixed period of 5 seconds. This is not a problem per-se but is not accuaret, instead the code should sleep for an interval which guarantees on next wakeupcommit_intervalwould have passed. Since time tracking is not accurate add 1 second to ensure next wake up would be aftercommit_interval.Signed-off-by Nikolay Borisov <[email protected]>--- fs/btrfs/disk-io.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 14 lines
fs/btrfs/disk-io.c static int transaction_kthread(void *arg) · line 1738
Quoted 10 lines
fs/btrfs/disk-io.c +1
-1@@ -1735,7 +1735,7 @@ static int transaction_kthread(void *arg)if (cur->state <TRANS_STATE_COMMIT_START && delta < fs_info->commit_interval) {spin_unlock (&fs_info->trans_lock); - delay =msecs_to_jiffies (5000 ); + delay =msecs_to_jiffies ((1 +delta) *1000 );This does not seem right. Delta is number of seconds since the transaction started, so we need to sleep for the remaining time. Which is
commit_interval- delta.
Doh, you are right. The correct line should be :
delay -= msecs_to_jiffies((1+delta) * 1000);
The + 1 is needed so that upon next wake up we are guaranteed next delta is >= commit_interval. Shall I send a fixed patch or are you going to modify it and merge it?
Quoted 1 line
Re: [PATCH 4/4] btrfs: Be smarter when sleeping in transaction_kthread
Quoted 32 lines · Nikolay Borisov
Quoted 14 lines · David Sterba
Quoted 13 lines · Nikolay Borisov
If
transaction_kthreadis woken up beforebtrfs_fs_info::commit_intervalseconds have elapsed it will sleep for a fixed period of 5 seconds. This is not a problem per-se but is not accuaret, instead the code should sleep for an interval which guarantees on next wakeupcommit_intervalwould have passed. Since time tracking is not accurate add 1 second to ensure next wake up would be aftercommit_interval.Signed-off-by Nikolay Borisov <[email protected]>--- fs/btrfs/disk-io.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)Quoted 14 lines
fs/btrfs/disk-io.c static int transaction_kthread(void *arg) · line 1738
Quoted 10 lines
fs/btrfs/disk-io.c +1
-1@@ -1735,7 +1735,7 @@ static int transaction_kthread(void *arg)if (cur->state <TRANS_STATE_COMMIT_START && delta < fs_info->commit_interval) {spin_unlock (&fs_info->trans_lock); - delay =msecs_to_jiffies (5000 ); + delay =msecs_to_jiffies ((1 +delta) *1000 );This does not seem right. Delta is number of seconds since the transaction started, so we need to sleep for the remaining time. Which is
commit_interval- delta.
Quoted 3 lines
Doh, you are right. The correct line should be :
delay -=
msecs_to_jiffies((1+delta) * 1000);
Correction it should be :
delay -= msecs_to_jiffies((delta-1) * 1000);
Quoted 7 lines
The + 1 is needed so that upon next wake up we are guaranteed next delta is >=
commit_interval. Shall I send a fixed patch or are you going to modify it and merge it?Quoted 1 line
PATCH v2 1/1 btrfs: Be smarter when sleeping in transaction_kthread
In reply to David SterbaIf transaction_kthread is woken up before btrfs_fs_info::commit_interval seconds have elapsed it will sleep for a fixed period of 5 seconds. This is not a problem per-se but is not accuaret, instead the code should sleep for an interval which guarantees on next wakeup commit_interval would have passed. Since time tracking is not precise substract 1 second from delta to ensure the delay we end up waiting will be longer than than the wake up period.
---
This one survived my testing and debugging confirmed that sometimes delta can indeed be 0 so utilising min is the right thing to do.
fs/btrfs/disk-io.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
fs/btrfs/disk-io.c +2 -1
@@ -1735,7 +1735,8 @@ static int transaction_kthread(void *arg)if (cur->state <TRANS_STATE_COMMIT_START && delta < fs_info->commit_interval) {spin_unlock (&fs_info->trans_lock); - delay =msecs_to_jiffies (5000 ); + delay -=msecs_to_jiffies ((delta-1 ) *1000 ); + delay =min (delay,msecs_to_jiffies (fs_info->commit_interval *1000 )); goto sleep; } transid = cur->transid;
--
2.25.1
Quoted 9 lines · Nikolay Borisov
Following 4 patches make
transaction_kthreadcode slitghly more user friendly. Namely, patch 1 convert from open-coded multiplicaiton by HZ to usingmsecs_to_jiffieshelper. Patch 2 relies on the observation that the running transaction is obtained undertrans_lockso an extra check can be removed. Patch 3/4 could possibly be squashed into 1 but the net effect is that the code is more intuitive when sleeping in case a lower interval thancommit_transhas elapsed.
Quoted 7 lines
This has survived full xfstest run.
Nikolay Borisov (4):
btrfs: Use helpers to convert from seconds to jiffies in
transaction_kthread
btrfs: Remove redundant check
btrfs: Record delta directly intransaction_kthread
1-3 added to misc-next, with some fixups.
Quoted 1 line
btrfs: Be smarter when sleeping in
transaction_kthread
There's a comment regarding correctness.